BUG: perfect prediction check in DiscreteModel only applies to binary? no offset

[josef-pkt]

while browsing and skimming some code

The perfect prediction check in DiscreteModel ignores offset (new enhancement) and code uses properties that apply only for binary models, Logit and Probit.
If we would check for perfect prediction in countmodels like Poisson, then we would also need to include exposure.

    def _check_perfect_pred(self, params, *args):
        endog = self.endog
        fittedvalues = self.cdf(np.dot(self.exog, params[:self.exog.shape[1]]))
        if (self.raise_on_perfect_prediction and
                np.allclose(fittedvalues - endog, 0)):
            msg = "Perfect separation detected, results not available"
            raise PerfectSeparationError(msg)

I guess currently this method is only called by binary models.

[josef-pkt]

simple fix is to replace the explicit computation by call to predict

[josef-pkt]

bump, this is still a bug, now that we have offset in discrete binary models